MinIO AIStor: The Unified Data Foundation for the NVIDIA AI Factory

The NVIDIA AI Factory delivers the full-stack blueprint for industrializing enterprise AI — from data preparation and training to real-time agentic inference. But end-to-end throughput depends on the foundational data plane. When storage, governance, and metadata operations are not engineered for scale, GPU utilization drops and operational overhead increases. This whitepaper explains how MinIO AIStor serves as that unified data foundation across the entire AI Factory pipeline. For inference, AIStor is designed for BlueField-4 DPU integration to offload KV cache management — reducing recomputation and sustaining tokens/sec as context length and concurrency rise — and provides an S3 plugin for NVIDIA NIXL enabling multi-tenant KV cache tiering across inference runtimes. For RAG pipelines, AIStor anchors the document-to-embedding-to-index loop as the durable system of record. For lakehouse architecture, AIStor's native Iceberg catalog eliminates catalog sprawl by embedding table metadata directly in the storage platform, and AIStor Views publishes curated datasets as queryable data products without data movement. For training, AIStor provides strict S3 consistency, RDMA-enabled data paths, and predictable high-throughput access to multimodal datasets — delivering up to 5x training throughput compared to S3 over HTTP.